Table of Contents
On June 8, 2026, Tim Cook took the stage at WWDC26. Everyone knew it would be his last time opening Apple’s developer conference as CEO — on September 1, he transitions to Executive Chairman, handing the role to John Ternus. The keynote carried that weight, even if Apple didn’t dramatize it.
TL;DR
WWDC26’s headline: Apple embedded Google Gemini into Siri, turning Apple Intelligence from a feature collection into a genuine cross-app AI layer. Cook exits in September, Ternus steps in, and Apple’s AI strategy is clearly mid-repositioning.
What Was Announced
Siri’s Rebuild
Siri is now powered by Google Gemini. The new version understands multi-turn conversation, pulls context across apps — ask it about your afternoon meeting with someone and it reads Mail, Calendar, and Messages to answer — and ships as its own standalone app while remaining integrated system-wide.
Craig Federighi’s framing: “Data is only used to execute your request.” Apple is betting its differentiation is privacy architecture, not model capability.
Apple Intelligence Updates
This generation is more practical than the conceptual promises of a year ago:
- Spatial Reframing: Adjust a photo’s composition and perspective after the fact
- Extend: AI-generated background expansion beyond a photo’s original edges
- Cross-app context: Smart reply suggestions reference both Mail and Messages
- Natural language Shortcuts: Describe what you want, skip the manual setup
iOS 27 and macOS Golden Gate
iOS 27 rolls back to iPhone 11 compatibility — unusually broad for Apple. Key AI additions: rebuilt Spotlight, AI-integrated Photos and Mail search, system-wide dictation with automatic punctuation.
macOS jumps from the Sequoia family to Golden Gate.
One side note: foldable iPhone references surfaced in iOS 27 beta. No announcement, but developers are already digging.
Cook’s Exit Read
Ternus didn’t appear on stage at all — notable if Apple were actively positioning him as heir apparent. This could mean Cook got a clean farewell, or that Ternus’s strategic direction isn’t ready for public articulation yet.
The Google Gemini integration is arguably Cook’s most consequential AI decision: embedding a longtime competitor’s model at the core of Siri requires CEO-level conviction. How Ternus evolves that relationship — or doesn’t — is the most interesting open question.
Technical Take
Gemini in Siri signals Apple’s model strategy: Apple is not trying to win on foundation model quality. Their bet is that privacy architecture + deep system integration is a more durable moat than building the best LLM.
Cross-app context is technically hard: Reading across Mail, Calendar, Messages requires fine-grained permission layers and carefully bounded data access. Apple’s on-device + Private Cloud Compute split determines where the real capability ceiling is.
The on-device scale problem: Mobile hardware limits model size, but user expectations of AI capability are rising fast. Apple’s answer is hybrid edge-cloud — privacy is both the shield and the constraint.
What to Watch
- Ternus’s first WWDC (likely WWDC27): different AI narrative?
- Gemini boundaries in Siri: which requests stay on-device vs. Gemini, and how will Apple communicate this?
- Foldable iPhone: beta code is there, announcement timeline unclear
- Apple Intelligence internationalization: most features are English-only; Chinese and other language timelines?
References
Answers come from this article only. Click any prompt below or open the chat at the bottom right.
🇺🇸 English
On June 8th, 2026, Tim Cook walked onto the WWDC stage one last time as Apple's CEO. Everybody in that room knew it. Come September first, he steps into the role of Executive Chairman and hands the keys to John Ternus. Apple didn't make a big theatrical moment out of it — but the weight was there, hanging over the whole keynote.
So let's get to the headline, because it's a big one. Apple just embedded Google's Gemini model into Siri. Yes — Google, a longtime rival — now sits at the core of Apple's assistant. And that single move transforms Apple Intelligence from a scattered bag of features into a real, cross-app AI layer that runs across your whole device.
Here's what the new Siri can actually do. It handles back-and-forth conversation, so it remembers what you just said. And it pulls context from across your apps. Picture this: you ask about your afternoon meeting with someone, and Siri quietly reads your Mail, your Calendar, and your Messages to give you a real answer. It also ships as its own standalone app now, while staying woven into the system everywhere.
Craig Federighi summed up Apple's pitch in one line: "Data is only used to execute your request." And that tells you exactly where Apple is placing its bet. They're not claiming they have the smartest model — they're claiming they have the most private architecture. Privacy is the product.
Now, the rest of the Apple Intelligence updates. These feel a lot more grounded than the dreamy promises from a year ago. There's Spatial Reframing, which lets you tweak a photo's composition and perspective after you've already taken it. There's Extend, where the AI generates new background to stretch a photo beyond its original edges. Smart replies now reference both Mail and Messages together. And you can build Shortcuts just by describing what you want in plain language — no more fiddling with manual setup.
On the operating system side: iOS 27 pushes compatibility all the way back to the iPhone 11, which is unusually generous for Apple. The AI additions include a rebuilt Spotlight, AI-powered search inside Photos and Mail, and system-wide dictation that punctuates for you automatically. And macOS jumps out of the Sequoia era into a new family called Golden Gate.
One fun little breadcrumb — references to a foldable iPhone showed up in the iOS 27 beta. Nothing official, no announcement, but developers are already digging through the code.
Now, let's read the tea leaves on Cook's exit, because it's genuinely interesting. John Ternus, the man taking over, didn't appear on stage at all. Not once. If Apple were actively grooming him in public as the heir apparent, you'd expect a moment. So this could mean one of two things: either Cook got a clean, uninterrupted farewell — his stage — or Ternus's strategic vision just isn't ready to be spoken out loud yet.
And that Gemini deal? That might be Cook's most consequential AI decision, period. Putting a direct competitor's model at the heart of Siri takes CEO-level conviction. It's a bold, almost uncomfortable choice. Which makes the biggest open question this: what does Ternus do with that relationship? Does he lean in, or pull back?
Let me dig into the technical stakes for a second, because there's real substance here. Apple's model strategy is now crystal clear — they are not trying to win the foundation-model race. They've decided that privacy plus deep system integration is a stronger, more durable moat than building the world's best language model. That's a philosophical bet, not just a product one.
But cross-app context is genuinely hard to pull off. Reading across Mail, Calendar, and Messages means building fine-grained permission layers and carefully fencing off what data the AI can touch. And Apple splits the work between on-device processing and something they call Private Cloud Compute — and where that line falls determines the real ceiling on how capable this thing can be.
Which points to the core tension: mobile hardware can only run so large a model, but people's expectations for AI keep climbing fast. Apple's answer is this hybrid, edge-plus-cloud approach. And here's the elegant, slightly painful irony — privacy is both their shield and their constraint. The very thing that protects you also limits how far they can push.
So what should you keep an eye on going forward? First, Ternus's first WWDC, probably WWDC27 — will the AI story sound different under him? Second, the exact boundaries of Gemini inside Siri — which requests stay on your device, which get sent to Google, and how honestly Apple explains that split. Third, that foldable iPhone lurking in the beta. And fourth, internationalization — most of these features are English-only right now, so the timeline for Chinese and other languages is a real open question.
Let me leave you with the three things that matter most. One: Apple has quietly conceded the model-quality race and is betting everything on privacy architecture as its moat — Gemini inside Siri is the loudest proof. Two: this was Tim Cook's farewell, and John Ternus's near-total absence from the stage is a signal worth watching, whichever way you read it. And three: the whole strategy lives or dies on that edge-cloud balance — privacy is Apple's greatest strength and its tightest leash, all at once.
Apple's AI story isn't finished. It's mid-pivot. And the most interesting chapter starts the day Ternus takes over.
🇹🇼 中文
2026 年的 WWDC,是 Tim Cook 以 CEO 身分主持的最後一場大會。台下給了他將近一分鐘掌聲,但真正的問題不在這場謝幕,而在後面——今年 9 月起,接棒的新 CEO John Ternus 得帶著蘋果,在 AI 時代重新證明「蘋果是定義者,不是追趕者」。
那麼,都 2026 年了、CEO 都要換人了,蘋果的 AI 到底做出了什麼?從投資人角度看,這場發布會其實有點平淡。三年半以來,這是外界第一次真的感覺到蘋果在「把 AI 塞進手機裡」,但端出來的東西並沒有超出市場預期。今天就順著矽谷 101 在現場的觀察,把幾件事講清楚:Apple Intelligence 失敗後蘋果內部經歷了什麼、重組後的領導班底、五款自研模型,還有跟 Google 合建的隱私架構。
先說危機的源頭。2024 年 6 月,蘋果正式宣布 Apple Intelligence,但之後功能一延再延。這不只讓業界質疑它的 AI 研發能力,更把「拿還沒交付的功能當宣傳核心」這件事,變成了消費者虛假廣告訴訟、還有股東證券詐欺訴訟的引信。到了 2025 年,蘋果毫無疑問陷入一場 AI 危機。
而危機的根源,其實是蘋果引以為傲的「完美文化」在 AI 衝擊下失衡了。蘋果過去很成功,靠的是整合能力——一年憋一個大招,穩穩往前走。這套節奏過去很有效,但在 AI 時代可能不適合。這個時代大概每週就該有一個 release,你必須激進。如果還像以前一年憋一個大招,那別人的 AI 都過了幾個世紀,你才過了一年,這就是問題。
接著是一場改寫權力版圖的祕密會議。根據彭博社報導,2025 年初那場關鍵會議,Cook 本人沒出席,由當時的 COO Jeff Williams 主持。蘋果總算意識到:Apple Intelligence 的問題不只是產品延期,而是組織、文化跟領導力的問題。
原本負責 AI 的 John Giannandrea,正在失去 Cook 跟高層的信任。蘋果需要一個人來收拾 Siri 跟 AI 的爛攤子——這個人是 Mike Rockwell,之前做 Vision Pro 跟 visionOS 的那位。Vision Pro 銷量不算成功,但在內部被視為工程難度極高的成果,Rockwell 是打過硬仗、帶過隊、做過複雜系統的人,而且他很早就認為 AI 會是蘋果的關鍵,也提過 Siri 的改造路線。
不過這裡出現了大公司常見的政治博弈。Rockwell 原本以為自己會成為蘋果整體 AI 的負責人,取代 Giannandrea,直接向 Cook 匯報;但軟體工程老大 Craig Federighi 不同意——在他看來,AI 跟 Siri 最終應該屬於軟體工程體系。最後結果是:Rockwell 接下 Siri,但向 Federighi 匯報,而不是直接向 Cook;Giannandrea 則離開了蘋果。這背後其實是蘋果對一個核心問題的重新劃分——AI 到底是一個獨立的新中樞,還是軟體系統的一部分。
重組後的權力版圖大概是這樣:Tim Cook 從旁觀產品,變成親自下場介入 AI 路線;Craig Federighi 依然是軟體負責人,而 AI 成了未來幾年作業系統升級的中心;Mike Rockwell 接手並重組 Siri 團隊。還有一個關鍵角色,是蘋果從外部挖來的 AI 模型負責人 Amar Subramanya——他在 Google 待過 16 年,做過 Gemini、Bard 的工程副總裁,剛跳槽微軟才 5 個月就被蘋果挖走。今年蘋果跟 Gemini 的合作,相信他中間有不少話語權。至於 John Ternus,今年 9 月接替 Cook,他繼承的是一個「必須證明 AI 能力」的蘋果。所以今年的 WWDC,就是這整個新高層團隊的檢驗時刻。
好,接下來是重頭戲:五款自研模型。新一代的 Apple Foundation Models,簡稱 AFM,一共五款,兩顆跑在端側、三件套在雲端。發布會反覆強調的是:就算進了 AI 時代,蘋果端出來的也得是符合蘋果價值觀的技術。
端側兩顆。第一顆 AFM 3 Core,30 億參數。第二顆 AFM 3 Core Advanced,是 200 億參數的 MoE 混合專家架構——雖然總參數 200 億,但一次大概只激活 10 到 40 億,實際計算量跟 30 億模型差不多,剛好能在手機端撐住。它的核心優勢是語音,更擬人、更有表現力、可個性化,還有全系統聽寫這些多模態能力。代價是硬體門檻高:Advanced 只能跑在 iPhone 17 Pro 系列、搭 M4 的 iPhone Air 以上,以及至少 12GB 統一記憶體的 iPad 跟 Mac。這也是部分用戶吐槽的點,但端側 AI 對記憶體跟晶片的要求本來就高,某種程度上是沒辦法的事。
Advanced 能在端側跑 200 億參數,除了 MoE 稀疏激活,還用了一個有意思的技術:把一部分固定的基礎參數放到快閃記憶體 flash 裡。端側設備的 DRAM 本來就不夠,如果一次把所有參數載進記憶體,既浪費資源又耗電;蘋果的做法是常用參數放 flash,再根據需求動態載入不同參數,既少佔記憶體,也降低耗電。這正是蘋果「軟硬一體」的優勢,慢慢在大模型上發揮出來。
雲端三件套,參數細節都沒公布,只能從職能看。第一個 AFM Cloud,官方叫它「server-side workhorse」,雲端主力,端側撐不住時就接手,主打速度跟效率。第二個 AFM Cloud Pro,是最強的雲端模型,負責複雜推理跟 Agentic 的工具調用,用在最吃算力的場景,還擴展到了 Google Cloud 上專為 NVIDIA GPU 最佳化的環境。第三個是雲端圖像模型,負責圖像生成跟編輯。
這裡有個外界最關心的問題。兩年前蘋果跟 OpenAI 合作,本地用蘋果模型、雲上用 GPT,被批前沿能力不足;這次換成 Gemini,那模型到底是誰的?蘋果內部人士的回答很直接——五款模型全是蘋果自研的,但底層用到了 Gemini 的相關技術跟 Google Cloud 基礎設施,其中最重的 Cloud Pro 還擴展到 Google 的 NVIDIA GPU,並由蘋果的 Private Cloud Compute 隱私架構包住。換句話說,是蘋果主導模型,而不是把 Gemini 直接搬進蘋果平台。技術報告,蘋果說今年夏天稍晚會發,到時會分享更多細節跟 benchmark。
有了五款模型,靠什麼調度?關鍵字是「系統編排器」。它像演唱會的指揮,負責調度不同模型、組織不同功能來完成任務,底下掛著四個功能架構。第一個是 App Toolbox,這是讓 Siri 能「動手」的關鍵,透過 App Intents 框架,編排器知道每個 App 能做什麼,你叫它發訊息,它就調用 Messages 去執行。第二個是 Spotlight Index,把照片、郵件、備忘錄透過語意索引組織起來,讓 Siri 更懂你。第三個是螢幕感知,編排器看得到你當前螢幕的內容,能結合你正在讀的報導給答案。第四個是 World Knowledge,需要上網搜尋時,走蘋果跟 Google 合作的 Private Cloud Compute 生成回答。
發布會的 demo 老實說沒超出預期:新版 Siri 能結合個人情境搜郵件、照片、跨 App 做任務,能回答螢幕內容、也能上網。還有一個獨立的 Siri App,能回顧過往對話、透過 iCloud 私密同步。但大部分 demo 還是生成式 AI,寫作、修圖加多模態,沒有太多真正的 Agentic action。
不過從追趕角度看,這未必是壞事——它得先 catch up。對使用者來說,過去用 ChatGPT、用 Claude,手機卻一直沒跟上,要用 AI 得切到不同 App,記憶還整合不起來;如果 Siri 能把照片、郵件串起來,對用戶就是好事。這次比較驚豔的一個 demo,是透過密碼 App 自動幫你更新密碼——背後是 Apple Intelligence 跟 Safari 配合,自動導航到網站、登入、改密碼。這是很典型的 Agentic 場景,但也需要 Siri 跟開發者深度合作,而蘋果現在正處於搭建這種 Agentic 生態的初期。
講到開發者,就得講 App Intents,這套機制是用來取代舊的 SiriKit。它要求開發者把 App 裡的功能按結構化的模式宣告出來、開放給 Siri,Siri 就能透過工具調用連接每個 App。蘋果的推法是典型的「胡蘿蔔加大棒」。胡蘿蔔是:用戶量小於兩百萬的開發者,可以免費調用蘋果背後的 AFM 大模型。你想想,你自己去調 GPT-5.5 或 Claude 要付 token 費、還得自己搭一套 Agent Harness 去連 API,蘋果直接幫你搭好,你照規範寫程式碼就能免費用。大棒是:如果你不加入,未來越來越多人用 Siri 調 App 時,你就會被邊緣化——蘋果整套設備有數十億用戶,不在這個生態裡,很可能就拿不到 Siri 的流量。所以從開發者角度看,這其實很划算:降低成本,又多了一個 Siri 流量入口。
再來是隱私。要把 Consumer AI 推給全球上億使用者,安全是最重要的問題。這次蘋果跟 Google 花大力氣搭的隱私架構,叫 Private Cloud Compute,簡稱 PCC。先講結論:蘋果想建立一條「可驗證的計算鏈」——只有經蘋果批准、公開登記、完成硬體證明的伺服器,才能取得解密請求的能力;請求處理完,資料跟金鑰立刻銷毀;蘋果、Google、雲端管理員都沒有讀取用戶內容的常規介面。這就是發布會反覆強調的那句:連蘋果都不會知道你的資料。
為什麼這件事難?傳統雲端只在資料儲存時加密、傳輸時走 TLS,但 AI 推理的時候,GPU、CPU 必須讀取請求內容,資料通常得以可處理的形式出現在伺服器記憶體裡。PCC 想用硬體、密碼學跟軟體架構強制隔離這些資料。流程大概是這樣:端側先判斷請求要不要上雲;要上雲,設備會先驗證「對面是誰」,通過對方的密碼學證明才信任這個節點;在 TLS 之上再加一層保護,讓網路營運商、負載平衡器都看不到明文;請求只在 PCC 的信任邊界內解密,模型在受隔離的 CPU、GPU 環境處理;推理完,結果加密回傳、由設備解密;任務一結束,請求資料、臨時金鑰、運行狀態全部清除——蘋果把這叫做「無狀態計算」。
新版 PCC 最核心的變化,是把蘋果原本建立在 Apple Silicon 上的安全理念,遷移到 Google Cloud 的異構硬體上。這條信任鏈由三層硬體撐起。第一層 Intel TDX,在 CPU 側建立受硬體隔離的機密虛擬機,讓記憶體跟宿主機、其他租戶隔離。第二層 NVIDIA 機密計算,因為 Cloud Pro 跑在 GPU 上,光保護 CPU 記憶體不夠,參數會在 CPU、GPU 之間流動,所以要用 NVIDIA 的機密計算保護 GPU 推理,這裡結合的是 Intel CPU 加 NVIDIA Blackwell GPU,保護推理過程中的 data in use。第三層是 Google 的 Titan 晶片,作為硬體信任錨點,驗證平台身分跟啟動流程。再加上蘋果的軟體證明來驗證實際運行的程式碼,就串成一條完整的信任鏈:Titan 驗平台、Intel 隔離 CPU、NVIDIA 隔離 GPU、蘋果軟體證明。難點在於機密計算成本很高,業界過去也沒把這些整合成一套能全球規模運行的系統——這是蘋果第一次嘗試。這種投入,不是一家初創公司能做好的,需要 Google 這種 Infra 團隊,據了解 Google 確實抽調了不少人專門支援。
當然,砸了真金白銀在隱私上,監管還是硬骨頭。據紐約時報報導,歐盟委員會跟蘋果關於 AI 監管的爭議,導致新功能在歐盟延後發布。爭議點集中在《數位市場法案》DMA,它要求蘋果這類巨頭確保產品互操作性,意思是外部開發者能提供跟 Siri 競爭、可下載的 AI 助理。監管機構認為這對促進競爭很關鍵,蘋果則說這會帶來隱私跟安全隱患。這對蘋果非常重要——歐洲是僅次於美國的第二大市場,2025 年銷售額 1110 億美元;大中華區是第三大,同樣舉足輕重。
最後有個容易被低估的重點:兒童帳戶。這次發布會花了不少時間講它的大幅升級,涵蓋設備初始化、內容存取、通訊跟螢幕使用時間,秋季隨 iOS 27 系列推出。家長可以決定設備一開始有哪些 App,孩子想下載新 App、瀏覽新網頁、加新聯絡人,家長都會收到通知再決定批不批准。現場幾位科技行業媽媽的實測回饋很具體:最印象深刻的是 Screen Time,把孩子能碰的 App 分成三大場景,可以用溝通的方式一起商量娛樂跟學習時間,設定流程非常 easy,還能把時間決策權交給孩子,幫助培養自律。她們也給了建議——目前是到點才提醒,希望能提前 5 到 10 分鐘倒數,因為對小孩來說「戛然而止」最容易情緒崩潰。另外對青春期孩子的適配還不夠,有國情因素,比如家庭用微信而不是 Messages,而微信不開放管控。但對小小孩,這套機制能讓家長放心給他一台屬於自己的設備,而不是把自己的分享出去——這大概也會實打實幫到蘋果的銷量。
好,把這場 WWDC 收攏一下,三個核心要點。
第一,這是一次權力重組多過技術突破。從 Giannandrea 出走、Rockwell 接手 Siri、外部挖來 Amar Subramanya,到 Cook 親自下場、Ternus 即將接班——蘋果內部真正的變化,是把 AI 從邊緣拉回作業系統的中心。
第二,五款自研模型加系統編排器,路線很清楚:端側靠 MoE 加 flash 動態載入把 200 億參數塞進手機,雲端靠 Cloud Pro 打複雜推理,底層用了 Gemini 技術跟 Google Cloud,但主導權在蘋果手上。真章要等夏天的技術報告跟 benchmark。
第三,PCC 這條跨 Apple Silicon 跟 Google Cloud 的信任鏈,是這次最有分量的投入——Titan 驗平台、Intel 隔 CPU、NVIDIA 隔 GPU、蘋果做軟體證明,目標是「連蘋果都讀不到你的資料」。
但別忘了,大部分 demo 還是生成式 AI,真正的 Agentic 能力才剛起步。如果這一切只是一個開始,那麼真正屬於消費者 AI 的 iPhone Moment,或許還得再等一等。
Tags
Related Articles
WWDC 2026: Siri Rebuilt, Intel Macs Out, Liquid Glass Fixed
WWDC 2026's biggest moves: Siri rebuilt from scratch using Google Gemini, macOS Golden Gate ends Intel Mac support, and iOS 27 adds an opacity slider to fix Liquid Glass readability.
My Take on Apple's 2025 M4 Lineup
The M4 MacBook Air and Mac Studio are solid spec upgrades — 16GB standard RAM and massive memory bandwidth for local AI workloads. Apple Intelligence's Siri integration, however, remains frustratingly inconsistent. Hardware is ahead; software is still catching up.
Agentic Social Network: When AI Agents Start Socializing on Your Behalf
Starting from the DIY-setup craze that OpenClaw ignited, this piece unpacks the technical evolution behind the Agentic Social Network: the three-layer architecture of Model Router, Memory OS, and protocols, plus the key mechanisms driving Agents toward long-horizon companionship, proactive action, and self-evolution.